Global climate change is expected to be accompanied by changes in the composition of plant
functional types. Such changes are predicted to follow shifts in the percentage cover and abundance
of grass species, following the C3 and C4 photosynthetic pathways. These two groups differ in a
number of physiological, structural and biochemical aspects. It is important to measure these
characteristic properties because they affect ecosystem processes, such as nutrient cycling. High
spectral and spatial resolution remote sensing systems have been proven to offer data, which can be
used to accurately detect, classify and map plant species. The major challenge, however, is that the
spectral reflectance data obtained over many narrow contiguous channels (i.e. hyperspectral data)
represent multiple classes that are often mixed for a limited training-sample size. This is commonly
referred to as the Hughes phenomenon or “the curse of dimensionality”. In the context of
hyperspectral data analysis, the Hughes phenomenon often introduces a high degree of
multicollinearity, which is caused by the use of highly-correlated spectral predictors.
Multicollinearity is a prominent problem in processing hyperspectral data for vegetation
applications, due to similarities in the spectral reflectance properties of biophysical and biochemical
attributes. This study explored an innovative method to solve the problems associated with spectral
dimensionality and the related multicollinearity, by developing a user-defined inter-band correlation
filter function to resample hyperspectral data. The proposed resampling technique convolves the
spectral dependence information between a chosen band-centre and its shorter and longer
wavelength neighbours. The utility of the new resampling technique was assessed for discriminating
C3 (Festuca costata) and C4 (Themeda triandra and Rendlia altera) grasses and for predicting their
nutrient content (nitrogen, protein, moisture, and fibre), using partial least squares and random forest
regressions. In general, results obtained showed that the user-defined inter-band correlation filter
technique can mitigate the problem of multicollinearity in both classification and regression
analyses. Wavebands in the shortwave infrared region were found to be very important in regression
and classification analyses, using field spectra-only datasets. Next, the analyses were up-scaled from
field spectra to the new generation multispectral satellite, WorldView-2 imagery, which was
acquired for the Cathedral Peak region of the Drakensberg Mountains. The results obtained, showed
that the WV2 image data contain useful information for classifying the C3 and C4 grasses and for
predicting variability in their nitrogen and fibre concentrations. This study makes a contribution by
developing a user-defined inter-band correlation filter to resample hyperspectral data, and thereby
mitigating the high dimensionality and multicollinearity problems, in remote sensing applications
involving C3 and C4 grass species or communities. / Thesis (Ph.D.)-University of KwaZulu-Natal, Durban, 2013.
Identifer | oai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:ukzn/oai:http://researchspace.ukzn.ac.za:10413/9471 |
Date | 16 August 2013 |
Creators | Adjorlolo, Clement. |
Contributors | Mutanga, Onisimo., Cho, Moses A. |
Source Sets | South African National ETD Portal |
Language | en_ZA |
Detected Language | English |
Type | Thesis |
Page generated in 0.001 seconds